Market
GPU shortage 2026: what buyers face and how to secure GPUs
Capacity is tight, on-demand stock sells out and median prices rose over the past year. Here is what is behind it, and the practical steps that still get you GPUs on a date you can plan around.
FullGPU engineering
8 min read
The GPU shortage 2026 buyers keep running into is not a lack of GPUs on paper. New generations keep arriving, from NVIDIA's Blackwell Ultra B300 to AMD's MI355X. What is scarce is capacity you can actually rent: the GPU model you want, in the quantity you need, starting on a date you can plan around. This guide covers what that looks like from the buyer's side, why it is happening, and the practical moves that still get a team its hardware.
GPU shortage 2026: what buyers run into
If you have tried to add capacity this year, the pattern is familiar. Four things come up again and again.
- On-demand stock is often sold out. Popular configurations, especially 8-GPU servers built on recent GPUs, are frequently unavailable on hourly on-demand pages.
- Providers favor one-year contracts. Many providers now prefer to sell capacity on one-year terms, and on-demand inventory is often what remains after those commitments.
- Quotas and approvals. Hyperscalers typically require quota approvals before you can launch GPU instances, and they bill egress on top. Getting a quota raised adds its own lead time.
- Prices went up. Median on-demand prices for the H100, H200 and B200 all rose over the past year.
FullGPU's October 2026 survey of 38 public price sources puts numbers on that last point. The third column is simple arithmetic: if a price rises by a factor of 1 + x, a fixed budget buys 1 / (1 + x) as many hours.
| GPU | Median on-demand price per GPU-hour, change over the past year | What a budget set a year ago buys today |
|---|---|---|
| H100 | Up roughly 13% | About 11% fewer GPU-hours |
| H200 | Up roughly 27% | About 21% fewer GPU-hours |
| B200 | Up roughly 20% | About 17% fewer GPU-hours |
The practical consequence: a plan that assumed you could rent whatever you need, whenever you need it, at last year's hourly rate is now optimistic on all three counts.
Why GPU capacity is tight
There is no single cause, and you do not need a market forecast to plan around it. Three structural factors explain why the tightness persists.
Training and inference compete for the same fleets
Training runs reserve large blocks of GPUs for weeks or months at a time. Inference demand grows with every model that goes into production, and it runs around the clock. Both want the same parts, and recent GPUs with large memory are attractive for both: the 141 GB of an H200 or the 288 GB of a B300 serves a large model on fewer GPUs and trains it with less sharding.
New systems take a long time to deliver
An 8-GPU HGX server or a rack-scale system is not a shelf item. GPUs, high-bandwidth memory, network adapters, power and cooling all have to come together, then be installed, tested and connected. Each new generation (Hopper, Blackwell, Blackwell Ultra and next Rubin) restarts that ramp, and large buyers often commit to new systems well ahead of delivery.
Power and data-center space are finite
NVIDIA rates the H100 and H200 SXM at up to 700 W. The B200 and B300 are reported at around 1,000 W and 1,100 W, and AMD rates the MI355X at 1,400 W, liquid-cooled. Eight B200s are roughly 8 kW for the GPUs alone, before CPUs, memory, networking and cooling. Racks at that density need power delivery and cooling that many older data halls were not built for, and adding that capacity takes time.
None of these factors resolves in a quarter. The sensible planning assumption is that capacity stays tight through your next few purchasing cycles.
How billing models behave when supply is tight
The way you buy GPUs matters more in a tight market than in a loose one. The four common billing models in the industry behave very differently when demand outruns supply.
| Billing model | How it works | In a tight market | Fits best |
|---|---|---|---|
| On-demand, per second or per hour | Pay while the instance runs; release it and it is gone | Often sold out on popular configurations; the price can move between runs | Short experiments, when stock exists |
| Spot or interruptible | Discounted spare capacity that can be reclaimed at short notice | Spare capacity is, by definition, the first thing extra demand absorbs | Fault-tolerant batch jobs with frequent checkpoints |
| Reserved term, often one year or more | Commit to capacity for the whole term | The most reliable route to large allocations, at the cost of a long lock-in | Large, steady workloads with a known horizon |
| Monthly dedicated | Flat price per month for the same machine, renewed monthly | Once delivered, the hardware stays yours while you renew; a first order may wait for stock | Steady workloads that need certainty without a one-year commitment |
Hourly billing still wins when you need a GPU for a few days and can find one. A month has about 730 hours, so a monthly server pays off when you would otherwise run that GPU for a large share of them. Our comparison of monthly vs hourly GPU pricing works through the break-even.
Seven practical ways to secure GPUs in 2026
1. Rent by the month and keep the hardware
Holding capacity is easier than finding it. A monthly dedicated server stays yours as long as you renew it: no preemption, no re-queueing, no other tenant on the same GPU. This is the model we use at FullGPU: whole physical GPUs, billed monthly in advance at a flat price, with a one-month minimum and cancellation in any month with no fee.
2. Request capacity before you need it
Start procurement when a project is approved, not when the data pipeline is ready. Ask every provider for a delivery date in writing, and treat a vague "soon" as a no. For rack-scale systems such as the GB300 NVL72, plan in quarters rather than weeks.
3. Be flexible on the GPU model
Write down what you actually need: memory per GPU, total memory, GPU-to-GPU bandwidth and software stack. Then list every model that meets it. A team that will take any GPU with at least 141 GB and NVLink has far more options than a team that insists on one exact SKU. The table in the next section lists sensible substitutes.
4. Right-size memory
Over-asking makes a shortage worse for you. Weights take roughly parameters × bytes per parameter: about 140 GB for a 70B model in BF16, about 70 GB in FP8. In FP8 that model fits one H100 tightly or one H200 comfortably, instead of a pair of GPUs. Add 10–30% headroom for the KV cache, activations and runtime, then size. Our GPU memory calculator does the arithmetic, and the guide to how much VRAM an LLM needs explains the KV cache side.
5. Commit for several months when the workload is steady
Providers plan capacity around commitments. A 6- or 12-month term, or an order for several servers, is easier to schedule and gives you more to negotiate with than a series of single months. Keep the one-month option for anything experimental.
6. Watch stock signals
Prefer providers that publish stock per configuration. "Limited" means a configuration can run out at any time, and "in stock" today can be "limited" next week. When a configuration that meets your requirements shows as available, request it then rather than waiting for a better one. Where a provider offers restock alerts, use them.
7. Keep your stack portable
Package your environment in containers, pin driver, CUDA or ROCm and framework versions, and keep checkpoints somewhere you can copy them from quickly. A stack that moves between GPU types, or between NVIDIA and AMD, can take whatever capacity appears first.
If your first-choice GPU is sold out
Substitutes are rarely free: each one changes memory, bandwidth or software. This table spells out the trade, using GPUs from our own catalog as examples.
| Workload | First choice | Substitutes | What changes |
|---|---|---|---|
| 70B-class model, FP8 inference (about 70 GB of weights) | H200, 141 GB | B200, 180 GB; RTX PRO 6000, 96 GB; H100, 80 GB | The B200 costs more but adds memory and bandwidth; the RTX PRO 6000 has 1.6 TB/s against 4.8 TB/s; the H100 leaves little room for the KV cache |
| Single-node training or fine-tuning | HGX H200, 8 × 141 GB | HGX B200; HGX H100; 8× MI355X | The H100 has 80 GB per GPU, so more sharding; the MI355X has 288 GB per GPU but runs ROCm instead of CUDA |
| Very large or MoE models, long contexts | HGX B300, 8 × 288 GB | 8× MI355X, 8 × 288 GB; HGX B200, 8 × 180 GB | The MI355X matches the capacity on ROCm; the B200 has 1,440 GB in total instead of 2,304 GB |
| Image and video generation | RTX PRO 6000, 96 GB | L40S, 48 GB; RTX 5090, 32 GB | Less memory per GPU: smaller batches, or models that must be quantized or offloaded |
If you are considering the AMD route, our notes on running AI workloads on the AMD MI355X cover what ports easily and what does not.
How to evaluate a provider's real availability
"Available" on a landing page and a server you can log into on a known date are different things. Before you pay, check these points.
- Published stock per configuration. A provider that shows stock status for each configuration is telling you something you can check. "Contact sales" for everything tells you nothing.
- A written delivery date before payment. Not an estimate in a chat window: a date in the order confirmation.
- A refund if the date is missed. This is what turns a promise into a commitment.
- The exact configuration in writing. GPU model and count, memory per GPU, CPU, RAM, storage, network and location.
- Whole GPUs or slices. Ask whether the GPUs are dedicated to you or shared through vGPU, MIG partitions or time-sharing.
- Term and exit. Minimum term, renewal terms and cancellation notice.
This is how we sell every server: our pricing page shows stock status for each configuration, the configuration, price, location and delivery date are confirmed in writing before payment, and if we miss the date you can cancel for a full refund under our refund policy.
When the server arrives, verify it in the first five minutes:
# GPU model, count and memory as the driver sees them
nvidia-smi --query-gpu=index,name,memory.total --format=csv
# How the GPUs are connected (NV# means NVLink)
nvidia-smi topo -mBudgeting while prices rise
Rising prices are easier to absorb when you know the price in advance. Three habits help.
- Budget per GPU-month at your real utilization. An 8× H100 server at $16,900 per month works out to about $2.89 per GPU-hour if you use it around the clock. At 50% utilization, the effective cost per used GPU-hour doubles to about $5.79. If your utilization is low and you can find hourly stock, hourly can be the cheaper option.
- Price the substitutes before you need them. Decide in advance the most you will pay for a GPU-month on each acceptable model, so you can say yes within minutes when stock appears.
- Lock the price for the period you can predict. A flat monthly price is fixed for each paid month, so your largest line item stops moving with the hourly market. For longer horizons, ask about term pricing.
Our GPU cost calculator compares monthly and hourly totals for the hours you expect to run.
Frequently asked questions
Is there a GPU shortage in 2026?
Yes, in the practical sense: capacity is tight, on-demand stock for popular GPUs is often sold out, and many providers prefer one-year contracts. GPUs can still be rented, but usually with some lead time, on a term, or on a different model than your first choice.
How much have GPU rental prices risen?
FullGPU's October 2026 survey of 38 public price sources found that over the past year the median on-demand price per GPU-hour rose by roughly 13% for the H100, 27% for the H200 and 20% for the B200.
What is the fastest way to get GPUs during the shortage?
Be flexible on the GPU model, size memory precisely so you do not over-ask, and order from providers that publish stock and confirm a delivery date in writing. Once you have a server, renewing it is easier than finding new capacity.
Do I need a one-year contract to rent GPUs in 2026?
Not necessarily. Many providers prefer one-year terms, but monthly dedicated servers exist: at FullGPU the minimum term is one month and you can cancel in any month with no fee.
What should I check before paying for GPU capacity?
Ask for the exact configuration and a delivery date in writing, a full refund if that date is missed, and confirmation that the GPUs are whole and dedicated rather than shared slices. Then verify the hardware with nvidia-smi as soon as it is delivered.
Dedicated GPUs, billed monthly
Whole GPUs from $449/mo, root access, bandwidth included. Card or crypto.
See GPUs & pricing Talk to an engineer